A simple method of segmentation with subpixel accuracy

In order to identify boundaries with subpixel accuracy it is necessary to construct a continuous or at least high resolution model of the intensity function I inferred from the discrete values of the digitized image. An approach described by Haralick (1984) is to fit a cubic polynomial to the discrete image data in the region under consideration. In the paper an alternative method is described in which (i) the function I is modelled linearly rather than by a higher order polynomial; and (ii) region boundaries are identified with local isoluminance contours (i.e. lines of locally constant brightness) rather than zero crossings of the second derivative. This method is particularly simple, yet, as is demonstrated, yields useful results.<
>